MIMO radar main lobe interference suppression method based on semi-blind source separation

Abstract

The application discloses a MIMO radar main lobe interference suppression method based on semi-blind source separation and relates to the technical field of radar anti-interference. The application converts a complex convolution mixing problem into a linear blind source separation problem by performing frequency domain conversion and phase space reconstruction on a single channel signal after beam forming, and realizes signal separation by adopting adaptive parameter estimation and a FastICA algorithm, so that main lobe deceptive jamming can be effectively suppressed under the condition of low signal-to-noise ratio and small sample, and the target detection and tracking performance of the MIMO radar in a complex electromagnetic environment is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of radar anti-jamming technology, and in particular to a method for suppressing main lobe interference in MIMO radar based on semi-blind source separation. Background Technology

[0002] Suppression of main lobe deception interference in MIMO radar is one of the core technologies in radar anti-jamming. When a radar can effectively suppress main lobe deception interference, its target detection and tracking performance in complex electromagnetic environments can be significantly improved. Traditional main lobe interference suppression methods mainly rely on spatial filtering techniques, using beamforming to create nulls in the direction of interference to achieve interference suppression. However, when deception interference enters from the direction of the radar main lobe, it completely aliases with the target signal in the spatial domain, making it difficult for traditional spatial filtering methods to maintain the integrity of the target signal while suppressing interference.

[0003] In recent years, blind source separation technology has provided a new solution for radar anti-jamming. This method utilizes the statistical independence of target and jamming signals, achieving signal separation through independent component analysis (ICM). However, traditional blind source separation methods require the observation channel number to be no less than the source signal number, which is difficult to guarantee in practical radar systems. Especially after array beamforming, only a single-channel observation signal is usually obtained, making traditional multi-channel blind source separation methods unsuitable for direct application. Furthermore, while deep learning-based methods perform well in certain scenarios, they rely on a large number of labeled samples for training. In real battlefield environments, it is difficult to obtain sufficient samples to support network training, resulting in a significant decrease in recognition performance under small sample conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a MIMO radar main lobe interference suppression method based on semi-blind source separation, so as to effectively suppress main lobe deception interference under small sample conditions.

[0005] The technical solution adopted in this invention is: a MIMO radar main lobe interference suppression method based on semi-blind source separation, comprising the following steps:

[0006] Step S1: Perform frequency domain conversion on the single-channel mixed signal after beamforming of the MIMO radar receiving array. Use the known radar transmit waveform spectrum to preprocess the signal, convert the time-domain convolutional mixing model into a frequency-domain linear mixing model, and obtain the preprocessed frequency domain signal.

[0007] Step S2: Prepare key parameters for phase space reconstruction: using correlation integral-based... The method, or a data-driven approach based on fully convolutional networks (FCNs), adaptively estimates the optimal embedding dimension. and time delay ;

[0008] Step S3: Based on the attractor theory in chaotic dynamics, use the parameters obtained in step S2. The frequency domain sequence obtained in step S1 is reconstructed in phase space, and the single-channel frequency domain sequence is embedded into a high-dimensional phase space to reconstruct a multi-channel observation signal matrix.

[0009] Step S4: For the reconstructed multi-channel observation signal matrix, perform blind source separation using the Fast Independent Component Analysis (FastICA) algorithm, and separate multiple independent components by utilizing the statistical independence between signals;

[0010] Step S5: For the separated independent components, identify and extract the target signal component based on the time-frequency characteristics of the target component to achieve effective suppression of main lobe interference.

[0011] Specifically, step S1 includes the following sub-steps:

[0012] Step S11: Beamforming is performed on the output signal of the MIMO radar receiving array to obtain a single-channel hybrid time-domain signal. The discrete sequence is obtained by sampling the sequence and then a fast Fourier transform is performed to obtain its frequency domain representation. That is, the spectrum of the mixed signal, where Sampling time, Angular frequency;

[0013] Step S12: Mix the spectrum of the signal. Divide by radar transmitted waveform spectrum To eliminate the influence of the emission waveform, the convolutional mixture model is transformed into a linear mixture model:

[0014]

[0015] in, The first one of the preprocessed single-channel frequency domain sequences Each sample value, , and These represent the frequency domain channel responses of the target, deceptive interference, and clutter, respectively. This is equivalent noise.

[0016] Specifically, in step S2, a correlation integral-based approach is adopted. The method adaptively estimates the optimal embedding dimension. and time delay Specifically, it includes:

[0017] Define the correlation integral

[0018]

[0019] in, The number of FFT data points. This is a distance threshold used to determine whether the distance between two points in phase space falls within a given neighborhood. for The number of embedding points in the dimensional space, For unit step function, For supremum norm, , They are respectively;

[0020] Constructed based on correlation integrals Statistic:

[0021]

[0022] in, One-dimensional correlation integral m to the power of;

[0023] Calculate the preprocessed single-channel frequency domain sequence Standard deviation According to the formula Calculate the distance threshold set ,Right now ;in, A set of preset distance threshold multiples ;

[0024] For each candidate delay According to the distance threshold set Each distance threshold in Calculate the corresponding correlation integral to construct Statistical measure, then calculate all distance thresholds. of average of statistics Statistic and changes in average statistics Based on each candidate delay and Obtain the combined correlation metric ;

[0025] In all Find the first zero-crossing object in the list and its corresponding candidate delay. As the optimal delay ;

[0026] Find all The smallest one in the list, and the candidate delay corresponding to the minimum value is denoted as . Through formula Calculate the optimal embedding dimension .

[0027] Specifically, in step S2, a data-driven method based on fully convolutional networks adaptively estimates the optimal embedding dimension. and time delay Specifically, it includes:

[0028] A symmetrical encoder-decoder network structure is constructed based on multiple convolutional layers. All convolutional layers employ dilated convolutions with a fixed dilation factor of 2, and the number of convolutional layers is even. The first half of the convolutional layers acts as the encoder, and the second half as the decoder. The input data to the encoder-decoder network is a single-channel sequence that has undergone frequency domain preprocessing. In the encoder, the number of channels in the feature vector output by each convolutional layer gradually increases from the initial 32 to 128; in the decoder, the number of channels is gradually reduced from 128 to 32 in a symmetrical manner, ultimately outputting a 32-dimensional feature vector, the first 8 of which are used to predict the embedding dimension. The last 24 dimensions are used to estimate the delayed index. .

[0029] Specifically, the encoder-decoder network is trained end-to-end, based on correlation integrals. The optimal embedding dimension calculated by the method and time delay As a supervisory label, mean squared error loss is used for optimization.

[0030] Specifically, the encoder-decoder network consists of 8 convolutional layers, with the kernel sizes of each layer set as follows: 7×1, 7×1, 5×1, 5×1, 3×1, 5×1, 7×1, and 1×1.

[0031] Specifically, step S3 includes:

[0032] Step S31: Construct the phase space trajectory based on the estimation results of step S2;

[0033] Single-channel frequency domain sequence The time series observations are treated as a nonlinear dynamical system; the optimal embedding dimension is obtained based on the estimation. and time delay The phase space is reconstructed using the delay embedding method, and the frequency sampling points are indexed. Construct the corresponding phase points:

[0034]

[0035] in, Each phase point in It is A dimensional vector containing single-channel sequences. At any moment Dynamic state information;

[0036] Step S32: Combine all phase points to form a multi-channel observation matrix, constructing a matrix with dimension [missing information]. Multi-channel observation signal matrix :

[0037]

[0038] in, This represents the effective sequence length.

[0039] Specifically, step S4 includes:

[0040] Multi-channel observation signal matrix Perform FastICA algorithm initialization settings: using a matrix To observe the signal, a blind source separation model is established. ,in It is a mixed matrix. The source signal matrix to be separated contains independent components such as target, interference, and clutter;

[0041] The observed signal is preprocessed by centering and whitening before the unmixing vector is initialized. The objective function is to find the direction with the strongest non-Gaussianity by maximizing negative entropy.

[0042]

[0043] in, For mathematical expectation, It is not a quadratic function. It is a standard Gaussian random variable;

[0044] The unmixed vector is updated using the fixed-point iterative formula of FastICA. ,pass Obtain the separated independent components Each row vector corresponds to a source signal, including target echo, interference signal and clutter components.

[0045] Specifically, step S5 includes:

[0046] After obtaining multiple independent components separated by FastICA, the target signal is finally identified and extracted by analyzing the differences in the dynamic characteristics of each component in phase space. Since the target signal typically forms a regular, smooth attractor structure in phase space, while interference and clutter components exhibit disordered or chaotic characteristics, the target signal component can be extracted by identifying the component with the most regular attractor morphology, thus achieving effective suppression of main lobe interference. Specifically, after obtaining multiple independent components separated by FastICA, each independent component is input into an identification network (a small-sample prototype network method for radar active deception jamming identification) for discrimination. The output component with the highest probability of being classified as the target is taken as the target signal component, achieving target echo extraction and main lobe interference suppression.

[0047] The technical solution provided by this invention brings at least the following beneficial effects:

[0048] This invention transforms the complex convolutional mixing problem into a linear blind source separation problem by performing frequency domain conversion and phase space reconstruction on the single-channel signal after beamforming. It then uses adaptive parameter estimation and FastICA algorithm to achieve signal separation, which can effectively suppress main lobe deception interference under low signal-to-noise ratio and small sample conditions, and significantly improve the target detection and tracking performance of MIMO radar in complex electromagnetic environments. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 The invention presents a schematic diagram of a universal co-located MIMO radar system containing DRFM jamming and target signals, as well as a semi-blind source separation signal processing flow, demonstrating the signal processing flow in a scenario where the target and DRFM jammer coexist.

[0051] Figure 2 This is a schematic diagram of the processing procedure of the method of the present invention.

[0052] Figure 3 The diagram shows the framework of the fully convolutional network (FCN) proposed in this invention. The network consists of 8 convolutional layers and uses dilated convolution for automatic estimation of phase space reconstruction parameters.

[0053] Figure 4 This diagram illustrates the separation performance of the method of this invention under a single type of interference, showcasing the target-interference separation effect under different interference types (such as DFJ, ISFJ, ISRJ, RGJ, and VGJ), where the number of separation channels is shown. Delayed index Noise-to-interference ratio .

[0054] Figure 5 The waveform separation performance of three comparison methods (EVD-ICA, IMS, and VMD-ICA) is shown in the graphs. The separation effect of each method is compared under single interference (DFJ) conditions, and the signal-to-noise ratio is also considered. .

[0055] Figure 6 This is a waveform separation performance diagram of the method of the present invention under clutter and multiple interference coexistence conditions, including separation results in scenarios where the target and clutter coexist and mixed scenarios with multiple interference types (such as DFTJ, ISRJ, ISFJ).

[0056] Figure 7 Performance validation curves for fully convolutional networks, including training loss, validation loss, and embedding dimension. and delayed index The verification accuracy curve. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0058] This invention provides a method for suppressing main lobe interference in MIMO radar based on semi-blind source separation. It operates on the single-channel signal obtained after receiving beamforming and can effectively combat main lobe deception interference generated by jammers such as digital radio frequency memory, which is difficult to suppress by conventional spatial filtering.

[0059] In one embodiment, see Figure 1 and Figure 2 The specific implementation steps of the method of the present invention include:

[0060] Step S1: Perform frequency domain conversion and preprocessing on the single-channel mixed signal after beamforming of the MIMO radar receiving array, such as... Figure 1 As shown, it illustrates a general-purpose co-located MIMO radar system containing DRFM interference and target signals, as well as a flowchart of semi-blind source separation signal processing, including key steps such as beamforming and frequency domain conversion. The specific implementation is as follows:

[0061] Step S11: First, beamforming processing is performed on the output signal of the MIMO radar receiving array to obtain a single-channel signal. The signal contains target, interference, clutter, and noise components. The signal is then sampled to obtain a discrete sequence, and finally, a Fast Fourier Transform is performed to obtain its frequency domain representation. .

[0062] Step S12: Since the radar transmission waveform is known, its spectrum is denoted as... By dividing by To eliminate the influence of the transmitted waveform, the preprocessed frequency domain sequence is obtained:

[0063]

[0064] in, The first one of the preprocessed single-channel frequency domain sequences Each sample value.

[0065] The convolutional mixing problem is transformed into a linear mixing model, and the preprocessed single-channel frequency domain sequence is used. It includes target, interference, and clutter components, and the noise term is simplified, laying the foundation for subsequent blind source separation.

[0066] Step S2: Adaptively estimate the embedding dimension using the CC method or a fully convolutional network (FCN). and time delay This step prepares key parameters for phase space reconstruction, ensuring that the reconstructed multi-channel observations can effectively capture signal dynamics. It can be implemented using either the traditional CC method or the data-driven FCN method, as detailed below:

[0067] Step S21 (CC method): Employ the CC method based on association integrals to simultaneously estimate the embedding dimension. and time delay .

[0068] In this invention, by finding... Reaching the first local minimum or zero point As the optimal delay, and through Follow The change tends to stabilize, thus determining the minimum embedding dimension. The specific implementation is as follows: First, calculate the standard deviation of the frequency domain sequence to set a distance threshold; then calculate the sequence... Standard deviation ,according to Calculate the distance threshold set ,Right now ;in, A set of preset distance threshold multiples Then for each candidate delay Calculate its average Statistics and Changes in average statistics And thus obtain the combined correlation metric. Ultimately, the selection was made. The first zero-crossing point is used as the optimal delay , and select Delay corresponding to the global minimum point Through formula The optimal embedding dimension is calculated.

[0069] Step S22 (FCN method): To improve computational efficiency, a fully convolutional network (FCN) can be used for parameter estimation. For example... Figure 2 The diagram illustrates the FCN framework proposed in this invention. This network consists of eight convolutional layers and employs dilated convolutions for automatic estimation of phase space reconstruction parameters. The network input is a single-channel frequency domain sequence, and the output is the embedding dimension. and delayed index The estimated values ​​are obtained. During training, the parameters generated by the CC method are used as ground truth, and the network weights are optimized by minimizing the mean squared error loss.

[0070] Step S3: Reconstruct the phase space based on attractor theory to generate a multi-channel observation signal matrix. Utilizing the principles of chaotic dynamics, the single-channel frequency domain sequence is embedded into a high-dimensional phase space to simulate multi-channel observation. The specific implementation is as follows:

[0071] Step S31: Based on Takens' embedding theorem, construct the phase space trajectory using the parameters obtained in step S2. (The text repeats itself here, so the translation only includes the first instance.) and The phase space is reconstructed using the delay embedding method, and the frequency sampling points are indexed. Construct the corresponding phase points:

[0072]

[0073] Step S32: Combine all phase points to form a multi-channel observation matrix. The multi-channel observation matrix is ​​obtained. :

[0074]

[0075] The multi-channel matrix has a dimension of ,in It is the effective sequence length. This step represents the number of FFT data points. This step provides spatial diversity for subsequent separation, overcoming the limitations of single-channel observation.

[0076] Step S4: Use the FastICA algorithm for blind source separation and extract independent components. Input the reconstructed multi-channel observations into the FastICA algorithm, and use statistical independence to separate signal components. Implementation details are as follows:

[0077] The FastICA algorithm, based on the principle of maximizing negative entropy, performs independent component analysis on the observation matrix. The algorithm first centers and whitens the data to eliminate correlation; then iteratively optimizes the separation matrix, extracting independent components by maximizing non-Gaussianity. The separated components are denoted as... .

[0078] Step S5: Identify and extract the target signal component from the separated components. Based on the differences in the dynamic characteristics of each independent component in the reconstructed phase space, the target signal is extracted from the separated independent components to achieve interference suppression. The specific implementation is as follows:

[0079] First, the individual components separated from FastICA were... This is mapped onto the reconstructed phase space. The target signal component typically forms a regular, smooth attractor structure in the phase space, while interference and clutter components exhibit disordered or chaotic characteristics. By comparing the attractor regularity of each component, the component with the most regular attractor shape is identified as the target signal.

[0080] To improve recognition efficiency and automation, a pre-trained recognition network can be used to classify the separated components. The network learns the dynamic characteristics of the target component and outputs its probability. Interference components, due to their complex morphology, are easily distinguishable. Finally, the target component identified from the independent components is directly used as the target signal after interference suppression. Output. To quantitatively evaluate the extraction effect, waveform similarity can be used. Verification required:

[0081]

[0082] in, It is the target signal. It is the transmitted waveform.

[0083] In the specific implementation process, a simulation verification platform was built. The hardware platform consisted of a PC equipped with an Intel i7-7700 CPU, an NVIDIA GTX1060 GPU, and 32 GB of RAM. The software platform used MATLAB for simulation experiments. Key parameter settings are shown in Table 1, including carrier frequency, bandwidth, and sampling rate. These parameters were set based on typical radar scenarios to verify the performance of the method.

[0084] Table 1 Parameter Settings

[0085]

[0086] Simulation results demonstrate that this method can effectively separate the target under various conditions. Figure 4 As shown, it demonstrates the separation performance of the proposed method under single interference conditions, including interference types such as DFJ, ISFJ, ISRJ, RGJ, and VGJ, and the number of separated channels. Delayed index Noise-to-interference ratio As can be seen, the target component is clearly separated. For example... Figure 5 As shown, the waveform separation performance of three comparative methods (EVD-ICA, IMS, and VMD-ICA) under single interference (DFJ) conditions was compared, with a signal-to-noise ratio (SNR) of -10dB, highlighting the superiority of this method. Figure 6 As shown, it demonstrates the separation results of the proposed method under clutter and multiple interference conditions, including scenarios where the target and clutter coexist, as well as mixed scenarios with multiple interference types, proving the robustness of the method. Figure 7 As shown, it provides performance validation curves for fully convolutional networks, including training loss, validation loss, and embedding dimension. and delayed index The validation accuracy shows that FCN can converge quickly and accurately estimate parameters.

[0087] The present invention achieves effective suppression of MIMO radar main lobe deception interference through the above steps, and the method is both efficient and robust.

[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0089] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A MIMO radar main lobe interference suppression method based on semi-blind source separation, characterized in that, Includes the following steps: Step S1: Perform frequency domain conversion on the single-channel mixed signal after beamforming of the MIMO radar receiving array, preprocess the signal using the radar transmitted waveform spectrum, and convert the time-domain convolutional mixing model into a frequency-domain linear mixing model to obtain the preprocessed frequency domain sequence. Step S2: Prepare key parameters for phase space reconstruction: using correlation integral-based... The method, or a data-driven method based on fully convolutional networks, adaptively estimates the optimal embedding dimension. and time delay ; Step S3: Use the parameters obtained in step S2 The frequency domain sequence obtained in step S1 is reconstructed in phase space, and the single-channel frequency domain sequence is embedded into a high-dimensional phase space to reconstruct a multi-channel observation signal matrix. Step S4: For the reconstructed multi-channel observation signal matrix, perform blind source separation using the fast independent component analysis algorithm, and separate multiple independent components by utilizing the statistical independence between signals; Step S5: For the separated independent components, identify and extract the target signal component based on the time-frequency characteristics of the target component.

2. The method as described in claim 1, characterized in that, Step S1 includes: Step S11: Beamforming is performed on the output signal of the MIMO radar receiving array to obtain a single-channel hybrid time-domain signal. The discrete sequence is obtained by sampling the signal, and then a fast Fourier transform is performed to obtain the spectrum of the mixed signal. ,in Sampling time, Angular frequency; Step S12: Mix the spectrum of the signal. Divide by radar transmitted waveform spectrum Transform the convolutional mixture model into a linear mixture model: ； in, The first one of the preprocessed single-channel frequency domain sequences Each sample value, , and These represent the frequency domain channel responses of the target, deceptive interference, and clutter, respectively. This is equivalent noise.

3. The method as described in claim 1, characterized in that, In step S2, a correlation integral-based approach is adopted. The method adaptively estimates the optimal embedding dimension. and time delay Specifically, it includes: Define the correlation integral ； in, The number of FFT data points. This is the distance threshold in the correlation integral. for The number of embedding points in the dimensional space, For unit step function, For supremum norm, , They are respectively; the The and the first Phase points Constructed based on correlation integrals Statistic: ； in, One-dimensional correlation integral m to the power of; Calculate the preprocessed single-channel frequency domain sequence Standard deviation According to the formula Calculate the distance threshold set ,Right now ;in, A set of preset distance threshold multiples ; For each candidate delay According to the distance threshold set Each distance threshold in Calculate the corresponding correlation integral to construct Statistical measure, then calculate all distance thresholds. of average of statistics Statistic and changes in average statistics Based on each candidate delay and Obtain the combined correlation metric ; In all Find the first zero-crossing object in the list and its corresponding candidate delay. As the optimal delay ; Find all The smallest one in the list, and the candidate delay corresponding to the minimum value is denoted as . Through formula Calculate the optimal embedding dimension .

4. The method as described in claim 1, characterized in that, In step S2, the optimal embedding dimension is adaptively estimated using a data-driven method based on a fully convolutional network. and time delay Specifically, it includes: A symmetrical encoder-decoder network structure is constructed based on multiple convolutional layers. All convolutional layers employ dilated convolutions with a fixed dilation factor of 2, and the number of convolutional layers is even. The first half of the convolutional layers acts as the encoder, and the second half as the decoder. The input data to the encoder-decoder network is a single-channel sequence that has undergone frequency domain preprocessing. In the encoder, the number of channels in the feature vector output by each convolutional layer gradually increases from the initial 32 to 128; in the decoder, the number of channels is gradually reduced from 128 to 32 in a symmetrical manner, ultimately outputting a 32-dimensional feature vector, the first 8 of which are used to predict the embedding dimension. The last 24 dimensions are used to estimate the delayed index. .

5. The method as described in claim 4, characterized in that, The encoder-decoder network is trained end-to-end, based on correlation integrals. The optimal embedding dimension calculated by the method and time delay As a supervisory label, mean squared error loss is used for optimization.

6. The method as described in claim 4 or 5, characterized in that, The encoder-decoder network consists of 8 convolutional layers, with the kernel sizes of each layer set as follows: 7×1, 7×1, 5×1, 5×1, 3×1, 5×1, 7×1, and 1×1.

7. The method as described in claim 1, characterized in that, Step S3 includes: Step S31: Construct the phase space trajectory based on the estimation results of step S2; Preprocessed single-channel frequency domain sequence The time series observations are treated as a nonlinear dynamical system; the optimal embedding dimension is obtained based on the estimation. and time delay The phase space is reconstructed using the delay embedding method, and the frequency sampling points are indexed. Construct the corresponding phase points: ； in, Each phase point in It is A dimensional vector containing single-channel sequences. At any moment Dynamic state information; Step S32: Combine all phase points to form a multi-channel observation matrix, constructing a matrix with dimension [missing information]. Multi-channel observation signal matrix : ； in, This represents the effective sequence length.

8. The method as described in claim 7, characterized in that, Step S4 includes: Multi-channel observation signal matrix Perform FastICA algorithm initialization settings: using a matrix To observe the signal, a blind source separation model is established. ,in It is a mixed matrix. The source signal matrix to be separated; The observed signal is preprocessed by centering and whitening before the unmixing vector is initialized. The objective function is to find the direction with the strongest non-Gaussianity by maximizing negative entropy. ； in, For mathematical expectation, It is not a quadratic function. It is a standard Gaussian random variable; The unmixed vector is updated using the fixed-point iterative formula of FastICA. ,pass Obtain the separated independent components Each row vector corresponds to a source signal, including target echo, interference signal and clutter components.

Images